agentfile

Find what is wrong with the AI agent configuration your repository already has.

Your team uses Claude Code, Copilot, Cursor, Codex — each with its own instruction file, in its own format, in its own place. Nobody reads all of them at once, so they drift, contradict each other, and quietly cost context in every session.

agentfile reads them as they are and tells you what is wrong.

npx @agentfile/cli doctor

No setup. No file to adopt first. Nothing written to disk. Nothing it finds is ever executed — hooks, skills, commands and MCP configuration are read as text, and a clean result says "no pattern matched", never "this is safe".

---

The problem

The same rule, maintained in four places, drifting apart:

AGENTS.md                        ← the one that is current
CLAUDE.md                        ← a copy, edited last month
.github/copilot-instructions.md  ← a copy, probably out of date
.cursor/rules/main.mdc           ← different format, different rules

Symlinks solve the copying. They do not tell you that two of these disagree about the package manager, that a hook points at a script nobody committed, that a .mcp.json server will silently fail to load, or that 10,000 tokens load into every session before anyone types a word.

---

What it does

Nine questions, one command each. Run them in this order the first time:

npx @agentfile/cli doctor              # what is here, and what is wrong with it
npx @agentfile/cli context src/api.ts  # which configuration applies here, and why
npx @agentfile/cli audit               # what a hook, skill or MCP server could do
npx @agentfile/cli adopt               # plan a single source of truth

Then, once it is part of the build:

npx @agentfile/cli check               # fast enough for a pre-commit hook
npx @agentfile/cli validate --strict   # every layer, in CI
npx @agentfile/cli lint                # drifted copies, duplication, context cost
npx @agentfile/cli compile --check     # generated files still match their source
npx @agentfile/cli eval                # did the agent actually comply

Everything is deterministic: no model, no network, same tree in, same findings out. Every finding carries a stable code, a location, why it matters, and what to do about it.

---

How it works

agentfile reads whatever is there, normalises it into one representation with full provenance, and answers every question from that:

AGENTS.md, CLAUDE.md, .cursorrules          ┐
.claude/{rules,skills,agents,commands}/     │
.cursor/{rules,commands}/                   ├─→  one normalized model  ─→  doctor
.github/copilot-instructions.md             │      every node knows          check
.github/instructions/, .agents/skills/      │      its file, line,           lint
.claude/settings.json, .mcp.json            ┘      platform and scope        audit
                                                                             context
                                                                             compile

Because there is one resolver, context and compile cannot disagree about what applies where.

If you would rather keep one source and generate the rest, adopt plans that and compile maintains it. That is a choice, not a prerequisite.

---

Getting started

Point it at any repository:

npx @agentfile/cli doctor

It needs nothing from you. If it finds nothing, it says what it looked for.

---

Installation

npm install --save-dev @agentfile/cli

Add to package.json:

{
  "scripts": {
    "ai:check":    "agentfile check",
    "ai:validate": "agentfile validate --strict"
  }
}

Requires Node.js >=22.0.0.

The local dashboard is a separate install, because it ships an HTTP server most projects never want: npm install --save-dev @agentfile/ui.

---

Documentation

• Diagnostic codes — every AGFxxx, what it means, and how to configure it
• What is stable — what CI and editors may depend on
• Moving to v2 — nothing you have breaks; here is what is new
• Security — what agentfile promises about execution, and what it does not promise about safety
• Contributing — the rules the code follows, and why

---

Commands

npx @agentfile/cli doctor

Analyses the AI agent configuration your repository already has — no setup required, nothing written to disk.

npx @agentfile/cli doctor
npx @agentfile/cli doctor --verbose      # list every file found
npx @agentfile/cli doctor --format json  # machine-readable, for CI

It reads AGENTS.md, CLAUDE.md, .claude/rules/, .claude/skills/, .claude/agents/, .claude/commands/, .cursor/rules/, .cursor/commands/, .github/copilot-instructions.md, .github/instructions/, and .mcp.json, then reports:

• what configuration exists, per platform, and where it lives
• how much context loads into every session (an estimate, clearly labelled)
• rules duplicated across platforms — the drift the rest of agentfile prevents
• skills whose description is too thin for an agent to route on
• misconfigurations such as an MCP server that will silently fail to load, or an instruction file importing a path that does not exist

doctor runs no model, makes no network calls, and never executes a hook, script, or MCP command it finds. It exits non-zero on errors, so it can gate CI.

npx @agentfile/cli adopt

Proposes a single source of truth for the configuration you already have, and shows the plan before touching anything.

npx @agentfile/cli adopt                  # plan only — writes nothing
npx @agentfile/cli adopt --apply          # carry it out, after confirming
npx @agentfile/cli adopt --source claude  # consolidate into CLAUDE.md instead

Adoption happens in two phases, and the order is not cosmetic. A compiler never carries a target's own file into that target, so generating CLAUDE.md while CLAUDE.md still holds text nothing else has would lose that text. So:

1. Consolidate. Everything every platform says is gathered into one file — AGENTS.md by default, because it is the cross-tool standard — which stays hand-written. Bodies are appended whole under a heading naming where they came from: nothing is rewritten, reordered, or summarised, and a file the source already says everything from is skipped rather than copied again.
2. Generate. The other platforms' files become compiler output of that source.

Nothing is written without --apply, and --apply confirms first. A hand-written file is overwritten only once its own text is already in the source — anything else is still refused, exactly as compile refuses it. Skills, subagents, commands, hooks, MCP servers and permission rules are left alone, and the plan says so rather than leaving you to notice.

Asking compile to do both phases at once is reported as AGF205: with two hand-written targets each becomes the other's source, and under --force their contents swap.

npx @agentfile/cli rule [code]

What a diagnostic code means, from the same registry that produces it.

npx @agentfile/cli rule            # every code, grouped by band
npx @agentfile/cli rule AGF302     # one code in full

npx @agentfile/cli context <path>

What configuration actually applies to a file — in load order, with the reason for each.

npx @agentfile/cli context src/api/handler.ts
npx @agentfile/cli context src/api/handler.ts --excluded   # and why the rest did not

When an agent behaves unexpectedly, the question is never "what does the configuration say" — it is which of these nine files reached this request, and which one won. This answers that: every instruction in load order with its platform and the reason it matched, the rules and skills available there, and the context cost at that path, separating what loads in every session from what is specific to the path.

npx @agentfile/cli explain <target>

The inverse question: where does this piece of configuration come from, and when does it apply?

npx @agentfile/cli explain .cursor/rules/api.mdc
npx @agentfile/cli explain deploy                            # a skill by name
npx @agentfile/cli explain "use pnpm" --at src/api/handler.ts

A target can be a file path, a skill or subagent name, or part of a rule's text. With --at, it answers whether the rule applies to that file, why or why not, and what outranks it there. It also names the other files declaring the same thing.

npx @agentfile/cli check

Fast deterministic validation, built for pre-commit hooks and editors. A full run takes around 140 ms including Node startup.

npx @agentfile/cli check
npx @agentfile/cli check --strict        # warnings fail the run
npx @agentfile/cli check --format json

Runs the structural and resolution checks: files that will not parse, references that point at nothing, the same rule maintained in several places, glob-scoped rules that match no file, and rules whose scope differs between platforms. No network, no model, nothing executed.

Silencing a finding you have already decided about

A finding you have reviewed and accepted is silenced with a comment in the file itself, in whatever comment syntax that file already uses:

<!-- agentfile-disable-next-line AGF302 mirrored deliberately for the audit trail -->

agentfile-disable-next-line covers the following line, agentfile-disable-line its own line, and agentfile-disable the whole file. Name no codes and it silences everything on that line.

Two things keep this from becoming a way to hide problems. Suppressed findings are counted, not discarded — every command says how many it silenced, and --format json carries each one with the directive responsible, so --no-suppressions is never needed to find out what a repository has chosen not to see. And a directive that silences nothing is reported as AGF005, so a suppression cannot quietly outlive the problem it was written for.

npx @agentfile/cli lint

Quality analysis — the things that are not wrong but are costing you.

npx @agentfile/cli lint
npx @agentfile/cli lint --budget 2000       # tighten the context budget
npx @agentfile/cli lint --similarity 0.75   # loosen near-duplicate detection

Finds copies of a rule that have drifted apart — exact comparison goes quiet at exactly the moment someone edits one copy and not the others — and measures always-loaded context against a budget, naming the largest contributors. Similarity is measured on words, not meaning, and the output says so.

For skills, it reports descriptions too thin for an agent to route on, two skills an agent has no basis to choose between, bodies past the size the specification recommends, and frontmatter that will not survive being shared through claude.ai or the Skills API.

npx @agentfile/cli validate

Strict validation across every layer. Designed for CI.

npx @agentfile/cli validate
npx @agentfile/cli validate --target claude    # what would compiling to Claude Code lose?
npx @agentfile/cli validate --target all
npx @agentfile/cli validate --strict
npx @agentfile/cli validate --list-rules       # print the rule set

With --target, it checks the features your configuration uses against what that platform actually supports, and every finding cites the platform's own documentation. Without --target it says compatibility was not checked, rather than assuming a target and failing your build over it.

It also validates every SKILL.md against the Agent Skills specification — agentfile validates against that standard rather than inventing a replacement — and statically inspects the scripts a skill bundles against a documented set of risk patterns. Nothing found in the repository is ever executed, and a clean result says "no pattern matched", never "this is safe".

ai/contract.yaml is still validated first and reported exactly as before, and a schema failure is still an immediate exit 1.

npx @agentfile/cli audit

Security and trust analysis of hooks, skills, MCP servers, and permission rules.

npx @agentfile/cli audit
npx @agentfile/cli audit --all      # include informational findings
npx @agentfile/cli audit --strict   # treat warnings as errors

Reads everything as text and executes nothing. Reports risky hook commands, unpinned MCP packages, committed credentials, permission rules that do not grant what they appear to, and prompt-injection indicators — each finding with the reason and, where it applies, the platform documentation that backs it. The output names every surface analysed and every file it could not read, and a clean result says what it means: no pattern matched, not "this is safe".

npx @agentfile/cli compile

Compiles the normalized configuration into native target files.

npx @agentfile/cli compile --target claude cursor    # from whatever your source of truth is
npx @agentfile/cli compile --target agents-md --check  # CI: exit 1 on drift, writes nothing

Whatever the repository maintains — AGENTS.md, CLAUDE.md, an agentfile contract — becomes the input; the requested targets get their documented file shapes (AGENTS.md files, CLAUDE.md plus .claude/rules, Copilot instruction files, Cursor .mdc rules). What a target cannot express is reported (AGF201–AGF203) instead of silently dropped, output is deterministic and marker-stamped, and a file agentfile does not own is never overwritten without --force.

npx @agentfile/cli eval

Behavioral evaluation: run an agent task in an isolated workspace, judge the result with deterministic assertions. See docs/evals.md.

npx @agentfile/cli eval --agent "claude -p {prompt}"
npx @agentfile/cli eval evals/button.eval.yaml --keep-workspace

Nothing executes in your working tree, no agent runs unless you name one, and results are cached against the repository state so unchanged evals are not re-run. Exit codes: 0 passed, 1 assertions failed, 2 harness error.

Generated-file utilities

These work with the manifest that both compile and the legacy sync write, so
they apply to either workflow.

npx @agentfile/cli diff

Checks generated files against .agentfile-manifest.json and exits non-zero when drift is detected.

npx @agentfile/cli diff
npx @agentfile/cli diff --files CLAUDE.md,.github/copilot-instructions.md

npx @agentfile/cli clean

Removes generated files that can be regenerated and updates manifest ownership records.

npx @agentfile/cli clean --dry-run
npx @agentfile/cli clean

npx @agentfile/cli rollback

Restores files from .agentfile-backup/.

npx @agentfile/cli rollback --list
npx @agentfile/cli rollback --tag migrate-1700000000000

Configuring agentfile

Everything below is also a flag, and a repository that never writes this file loses nothing. What the file buys is agreement: a pre-commit hook, a CI job and an editor cannot each spell out the same --budget 2000 --similarity 0.75, and a tool arguing for one source of truth should not need its own settings in three places.

agentfile.yaml, at the repository root, every key optional:

# Directory names to skip, added to the built-in list
ignore:
  - fixtures

# Per-code severity. `off` silences a code repository-wide
severity:
  AGF302: info
  AGF501: error
  AGF203: off

budget: 2000        # always-loaded context budget, in estimated tokens
similarity: 0.75    # near-duplicate threshold
targets: [claude, copilot]
maxWarnings: 0      # fail when warnings exceed this
suppressions: true  # honour agentfile-disable directives

A key nobody recognises is an error, not a shrug — a silently ignored sevrity: block is a setting the team believes is applied and is not, which is the failure agentfile exists to report. And a file that does not validate is reported and then ignored entirely: a half-applied configuration is worse than none, because the half that applied looks like the whole. A flag always beats the file.

Turning a code off in this file is a repository-wide decision, recorded in a committed file. It is a different thing from an agentfile-disable comment, which silences one finding at one line and is reported when it goes stale.

GitHub Action

Two lines in a workflow, rather than a command someone has to remember:

- uses: dennishavermans/agentfile@v1

The action runs check against the repository root, writes agentfile.sarif,
and fails the step when there are findings.

Uploading that SARIF is a separate step on purpose. It needs
security-events: write, and a job that asks for write access to your security
alerts should say so where you can see it, rather than acquire it as a side
effect of a step called "run the linter":

name: agentfile
on: [push, pull_request]

permissions:
  contents: read
  security-events: write

jobs:
  agentfile:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5
      - uses: dennishavermans/agentfile@v1
        id: agentfile
        with:
          fail-on-findings: false   # report first, gate once the count is down
      - uses: github/codeql-action/upload-sarif@v4
        if: always()
        with:
          sarif_file: ${{ steps.agentfile.outputs.sarif-file }}
          category: agentfile-check

if: always() matters once you turn gating on: without it, the upload is
skipped exactly when there is something to upload.

Inputs

input	default	what it does
command	check	check, validate, lint or audit — the four that produce findings against a file
root	.	directory to analyse
version	latest	version or dist-tag of @agentfile/cli; pin an exact version for a job that cannot change under you
sarif-file	agentfile.sarif	where to write SARIF; empty string skips it
fail-on-findings	true	whether findings fail the step
args	none	extra arguments, e.g. --strict or --max-warnings 5

Outputs

output	what it is
exit-code	0 clean, 1 findings, 2 the tool could not run
findings / errors / warnings	counts, for a later step to act on
sarif-file	path to the log, empty when SARIF was not requested

Running more than one command means more than one upload, and each needs its
own category (agentfile-check, agentfile-audit) or the second will close
the first one's alerts.

---

CI output

check, validate, lint and audit emit SARIF 2.1.0 with --format sarif, which GitHub code scanning reads. The action above is the packaged
version of this; by hand it is:

- run: npx @agentfile/cli check --format sarif > agentfile.sarif
  continue-on-error: true
- uses: github/codeql-action/upload-sarif@v4
  with:
    sarif_file: agentfile.sarif

Findings then appear as annotations on the pull request that introduced them instead of in a job log. Every code is declared with a documentation link, and findings are fingerprinted without their line number, so editing a file above a finding does not close the alert and open an identical new one.

--max-warnings <n> is the ratchet for working a warning count down: it fails the run when warnings exceed the ceiling, without --strict's claim that every warning is an error.

---

Legacy: the v1 contract workflow

The commands below generate files from an ai/contract.yaml that a repository
has to adopt first. They still work, are still tested, and nothing about their
output has changed — but they are no longer the recommended entry point. The v2
commands above read the configuration a repository already has and need no
adoption step, and agentfile adopt is the supported route from one to the
other.

npx @agentfile/cli init

Interactive setup. Scaffolds ai/contract.yaml, agent templates, .ai-agents.example, and a CI workflow. Safe to run in existing projects — never overwrites existing files.

npx @agentfile/cli migrate

Imports existing instruction files and creates a draft ai/contract.yaml.

npx @agentfile/cli migrate --from .github/copilot-instructions.md --from CLAUDE.md
npx @agentfile/cli migrate --from CLAUDE.md --replace-policy archive
npx @agentfile/cli migrate --from CLAUDE.md --targets claude,copilot

Key options:

• --replace-policy keep|archive|delete
• --targets <ides>
• --exclude <ides>
• --dry-run
• --output <path>

npx @agentfile/cli watch

Watches ai/ for changes and automatically re-runs sync. Runs an initial sync on start.

npx @agentfile/cli watch

Solves the problem of manually running sync after every contract change.

npx @agentfile/cli sync

Reads your personal .ai-agents file and generates the corresponding instruction files.

npx @agentfile/cli sync             # generate files
npx @agentfile/cli sync --dry-run   # render without writing — used in CI

npx @agentfile/cli ui

Starts a local dashboard in your browser for inspecting your contract, previewing generated files, and monitoring drift.

npx @agentfile/cli ui                        # open dashboard on default port 4311
npx @agentfile/cli ui --port 3000            # custom port
npx @agentfile/cli ui --root ./packages/app  # inspect a specific directory

Needs @agentfile/ui installed alongside the CLI (npm install --save-dev @agentfile/ui); it is not pulled in by default, because it ships an HTTP server and a built front end that most installs never use.

---

Your rules file

ai/contract.yaml is your single source of truth:

version: 1

project:
  name: My Project
  stack:
    - typescript
    - react
    - nextjs

rules:

  coding:
    - Prefer small composable functions
    - Avoid unnecessary abstractions

  architecture:
    - Follow feature-based folder structure
    - Avoid cross-feature imports

  testing:
    - Critical business logic must have unit tests

  naming:
    - Use descriptive variable names
    - Boolean variables must be prefixed with is, has, or should

---

Personal agent selection

Each developer creates a .ai-agents file (gitignored) listing the agents they personally use:

# .ai-agents — yours, not the team's
claude
cursor

A committed .ai-agents.example documents the available options for new joiners.

---

Skills

Skills define shared workflows that every agent understands. Define them once — each agent receives them in its own format.

version: 1

project:
  name: My Project
  stack: [typescript, react]

rules:
  coding:
    - Prefer small composable functions

skills:
  - name: create-component
    description: Creates a new React component with tests
    context:
      - Components live in /src/components/{feature}/
      - Always use TypeScript, never .jsx
    steps:
      - Create /src/components/{feature}/{Name}.tsx
      - Create /src/components/{feature}/{Name}.test.tsx
      - Export from /src/components/{feature}/index.ts
    expected_output: A typed component with a matching test and barrel export
    examples:
      - input: "create a UserCard component"
        output: "src/components/users/UserCard.tsx + test + index"

Each agent receives skills in its native format:

Agent	Output
Claude	Appended to CLAUDE.md as markdown sections
Cursor	One .mdc file per skill in .cursor/rules/skills/
Copilot	Compact inline bullets in copilot-instructions.md
AGENTS.md	Full markdown — also read natively by Codex and Windsurf

---

Supported agents

Agent	Generated file
claude	CLAUDE.md
copilot	.github/copilot-instructions.md
cursor	.cursor/rules/main.mdc + one .mdc per skill
agents-md	AGENTS.md — read natively by Codex and Windsurf

Adding a new agent requires only a new folder in ai/agents/. No generator changes, no config edits.

---

Per-folder overrides (monorepos)

Create an ai.override.yaml in any package directory to inject additional context into generated files for that package:

blocks:
  - section: Frontend Context
    content: |
      This package is a Next.js frontend.
      Prefer React Server Components by default.
      All API calls go through /lib/api — never call fetch directly.

Run npx @agentfile/cli sync from that directory and the override is injected automatically.

---

Adding a new agent

Drop a new folder into ai/agents/:

ai/agents/windsurf/
  config.yaml
  template.md

config.yaml declares the agent name and output path:

name: Windsurf
output: .windsurfrules
description: Windsurf AI rules file

template.md is your instruction template using ${token} syntax:

# ${project.name} Rules

## Coding
${rules.coding}

## Architecture
${rules.architecture}

## Skills
${skills}

---

Available tokens

Token	Output
${project.name}	Project name
${project.stack.join(', ')}	Stack as comma-separated string
${rules.coding}	Coding rules as markdown bullets
${rules.architecture}	Architecture rules as markdown bullets
${rules.testing}	Testing rules as markdown bullets
${rules.naming}	Naming rules as markdown bullets
${skills}	Skills rendered in agent-native format
${override}	Injected override blocks (if any)

---

CI integration

agentfile init generates a GitHub Actions workflow for you. It runs two checks on every change to ai/:

- name: Validate contract schema
  run: npx @agentfile/cli validate

- name: Dry-run generation
  run: npx @agentfile/cli sync --dry-run

The dry-run renders all templates without writing files — catching broken templates and invalid contracts before they reach developers.

---

.gitignore

Add the following to your .gitignore:

# agentfile — personal config and generated files
.ai-agents
CLAUDE.md
.github/copilot-instructions.md
.cursor/
.windsurfrules
AGENTS.md
ai.override.yaml
.agentfile-manifest.json
.agentfile-backup/

---

Programmatic usage

@agentfile/core is available as a standalone package for teams that want to integrate agentfile into their own tooling:

import { generate, validateContract } from '@agentfile/core'

// Validate only
const contract = validateContract({ contractPath: 'ai/contract.yaml' })

// Full generation
const result = generate({
  root:   process.cwd(),
  agents: ['claude', 'cursor'],
  dryRun: false
})

for (const r of result.results) {
  if (r.status === 'ok')      console.log(`Generated ${r.output}`)
  if (r.status === 'error')   console.error(r.error.message)
  if (r.status === 'skipped') console.warn(r.reason)
}

---

Packages

Package	Description
@agentfile/agentfile	Convenience wrapper — re-exports the full CLI
@agentfile/cli	CLI — init, migrate, sync, validate, watch, diff, clean, rollback, ui
@agentfile/core	Core engine — schema, loader, renderer, generator
@agentfile/ui	Local dashboard — interactive UI for managing the contract and inspecting generated files (beta, install with @beta tag)
VS Code extension	Sidebar, commands, and diagnostics directly in VS Code (beta, shows Preview badge on Marketplace)

---

Contributing

agentfile is in early development. Issues and pull requests are welcome,
including the ones that say a finding is wrong — that is the most useful kind of
report, and it has already changed the tool once.

npm install    # from the root
npm run build  # the CLI typechecks against core's built types, so build first
npm test
npm run lint

CONTRIBUTING.md has the rest: the rules the code follows and
why, how to add a diagnostic, and what makes a false-positive report
actionable.

---

Release process

Publishing happens in GitHub Actions, not on a laptop, so every package carries
npm provenance — a
signed statement of which commit and which workflow built it. For a tool that
tells people what is unsafe in their configuration, publishing without that
would be the wrong way round.

1. Update the version in packages/core, packages/cli and
   packages/agentfile, keeping the internal @agentfile/* dependency ranges
   in step. The release workflow refuses to publish if the tag and the manifests
   disagree.
2. Add the section to CHANGELOG.md. It becomes the release notes; the workflow
   fails if there is no section for the version being tagged.
3. Tag and push:

git tag v2.0.0-beta.1
git push origin v2.0.0-beta.1

That runs lint, build and the full test suite, publishes core → cli →
agentfile in that order, and creates the GitHub release. A version with a
hyphen in it is treated as a pre-release: it goes out under the next dist-tag
and is marked as a pre-release on GitHub, so npm install @agentfile/cli keeps
resolving to the last stable version.

workflow_dispatch runs the same thing as a dry run, which is worth doing once
before the first real tag.

Two tag namespaces live in this repository and only one of them releases
anything. A version tag is vX.Y.Z and publishes; the GitHub Action's moving
major tag is v1 and publishes nothing. The release workflow filters on
v*.*.* so that re-pointing v1 at a newer commit, which happens every time
the Action changes, cannot start a release.

One-time setup

• An NPM_TOKEN secret with publish rights, in a repository environment named
  release.
• Better, if you can: npm trusted
  publishing, which drops the token
  entirely and authenticates the workflow itself.

---

License

MIT